When I try to train autoencoder by importing images as numpy arrays the training proceeds quickly with the training loss at first epoch itself < 0 and the results are also decent.
But when I import same data through ImageDataGenerator the starting loss is around 32000 and as training proceeds it decreases very slowly and after 50 epochs it saturates at around 31000.
I used mse as loss function with Adam Optimiser. I tried different loss functions but the problem persists like Very high Value at start which saturates very quickly to significantly high value.
Any suggestions are welcomed. Thanks.
following is my code.
from convautoencoder import ConvAutoencoder
from tensorflow.keras.optimizers import Adam
import numpy as np
import cv2
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint
from tensorflow.config import experimental
from tensorflow.python.client import device_lib
devices = experimental.list_physical_devices('GPU')
experimental.set_memory_growth(devices[0], True)
EPOCHS = 5000
BS = 4
trainAug = ImageDataGenerator()
valAug = ImageDataGenerator()
# initialize the training generator
trainGen = trainAug.flow_from_directory(
config.TRAIN_PATH,
class_mode="input",
classes=None,
target_size=(64, 64),
color_mode="grayscale",
shuffle=True,
batch_size=BS)
# initialize the validation generator
valGen = valAug.flow_from_directory(
config.TRAIN_PATH,
class_mode="input",
classes=None,
target_size=(64, 64),
color_mode="grayscale",
shuffle=False,
batch_size=BS)
# initialize the testing generator
testGen = valAug.flow_from_directory(
config.TRAIN_PATH,
class_mode="input",
classes=None,
target_size=(64, 64),
color_mode="grayscale",
shuffle=False,
batch_size=BS)
mc = ModelCheckpoint('best_model_1.h5', monitor='val_loss', mode='min', save_best_only=True)
print("[INFO] building autoencoder...")
(encoder, decoder, autoencoder) = ConvAutoencoder.build(64, 64, 1)
opt = Adam(learning_rate= 0.0001, beta_1=0.9, beta_2=0.999, epsilon=1e-04, amsgrad=False)
autoencoder.compile(loss="hinge", optimizer=opt)
H = autoencoder.fit( trainGen, validation_data=valGen, epochs=EPOCHS, batch_size=BS ,callbacks=[ mc])
Ok. This was a silly mistake.
Adding rescale factor rescale=1. / 255 to imageDataGenerator solved the problem.
First of all I am not a programmer, but I am self-teaching me Deep Learning to undertake a real project with my own dataset. My situation can be broken down as follows:
I am trying to undertake a multiclass text classification project. I have a corpus with 1000 examples, each example with 4 possible labels(A1,A2,B1,B2) They are mutually exclusive. All the examples are in separate folders and separate .txt files.
After a lot of effort and some man tears I managed to put together this code:
import os
import string
import keras
import nltk
from nltk.corpus import stopwords
from nltk import word_tokenize
import re
import numpy as np
import tensorflow as tf
from numpy import array
from sklearn.model_selection import KFold
from numpy.random import seed
seed(1)
tf.random.set_seed(1)
root="D:/bananaCorpus"
train_dir=os.path.join(root,"train")
texts=[]
labels=[]
for label in ["A1","A2","B1","B2"]:
directory=os.path.join(train_dir,label)
for fname in os.listdir(directory):
if fname[-4:]==".txt":
f = open(os.path.join(directory, fname),encoding="cp1252")
texts.append(f.read())
f.close()
if label == 'A1':
labels.append(0)
elif label=="A2":
labels.append(1)
elif label=="B1":
labels.append(2)
else:
labels.append(3)
print(texts)
print(labels)
print("Corpus Length", len( root), "\n")
print("The total number of reviews in the train dataset is", len(texts),"\n")
stops = set(stopwords.words("english"))
print("The number of stopwords used in the beginning: ", len(stops),"\n")
print("The words removed from the corpus will be",stops,"\n")
## This adds new words or terms from words_to_add list to the stop_words
words_to_add=[]
[stops.append(w) for w in words_to_add]
##This removes the words or terms from the words_to_remove list,
##so that they are no longer included in stopwords
words_to_remove=["i","having"]
[stops.remove(w) for w in words_to_remove ]
texts=[[w.lower() for w in word_tokenize("".join(str(review))) if w not in stops and w not in string.punctuation and len(w)>2 and w.isalpha()]for review in texts ]
print("costumized stopwords: ", stops,"\n")
print("count of costumized stopwords",len(stops),"\n")
print("**********",texts,"\n")
#vectorization
#tokenizing the raw data
from keras.preprocessing.text import Tokenizer
from keras.preprocessing.sequence import pad_sequences
maxlen = 50
training_samples = 200
validation_samples = 10000
max_words = 10000
#delete?
tokens=keras.preprocessing.text.text_to_word_sequence(str(texts))
print("Sequence of tokens: ",tokens,"\n")
tokenizer = Tokenizer(num_words=max_words)
tokenizer.fit_on_texts(texts)
sequences = tokenizer.texts_to_sequences(texts)
print("Tokens:", sequences,"\n")
word_index = tokenizer.word_index
print("Unique tokens:",word_index,"\n")
print(' %s unique tokens in total.' % len(word_index,),"\n")
print("Unique tokens: ", word_index,"\n")
print("Dictionary of words and their count:", tokenizer.word_counts,"\n" )
print(" Number of docs/seqs used to fit the Tokenizer:", tokenizer.document_count,"\n")
print(tokenizer.word_index,"\n")
print("Dictionary of words and how many documents each appeared in:",tokenizer.word_docs,"\n")
data = pad_sequences(sequences, maxlen=maxlen, padding="post")
print("padded data","\n")
print(data)
#checking the encoding with a new document
text2="I like to study english in the morning and play games in the afternoon"
text2=[w.lower() for w in word_tokenize("".join(str(text2))) if w not in stops and w not in string.punctuation
and len(w)>2 and w.isalpha()]
sequences = tokenizer.texts_to_sequences([text2])
text2 = pad_sequences(sequences, maxlen=maxlen, padding="post")
print("padded text2","\n")
print(text2)
#cross-validation
labels = np.asarray(labels)
print('Shape of data tensor:', data.shape,"\n")
print('Shape of label tensor:', labels.shape,"\n")
print("labels",labels,"\n")
kf = KFold(n_splits=4, random_state=None, shuffle=True)
kf.get_n_splits(data)
print(kf)
KFold(n_splits=4, random_state=None, shuffle=True)
for train_index, test_index in kf.split(data):
print("TRAIN:", train_index, "TEST:", test_index)
X_train, X_test = data[train_index], data[test_index]
y_train, y_test = labels[train_index], labels[test_index]
#Pretrained embedding
glove_dir = 'D:\glove'
embeddings_index = {}
f = open(os.path.join(glove_dir, 'glove.6B.100d.txt'),encoding="utf-8")
for line in f:
values = line.split()
word = values[0]
coefs = np.asarray(values[1:], dtype='float32')
embeddings_index[word] = coefs
f.close()
print("Found %s words vectors fom GLOVE."% len(embeddings_index))
#Preparing the Glove word-embeddings matrix to pass to the embedding layer(max_words, embedding_dim)
embedding_dim = 100
embedding_matrix = np.zeros((max_words, embedding_dim))
for word, i in word_index.items():
if i < max_words:
embedding_vector = embeddings_index.get(word)
if embedding_vector is not None:
embedding_matrix[i] = embedding_vector
# define vocabulary size (largest integer value)
# define model
from keras.models import Sequential
from keras.layers import Embedding,Flatten,Dense
from keras import layers
from keras.layers.convolutional import Conv1D
from keras.layers.convolutional import MaxPooling1D
model = Sequential()
model.add(Embedding(max_words, embedding_dim, input_length=maxlen))#vocabulary size + the size of glove version +max len of input documents.
model.add(Conv1D(filters=32, kernel_size=8, activation='relu'))
model.add(MaxPooling1D(pool_size=2))
model.add(Flatten())
model.add(Dense(10, activation='relu'))
model.add(Dense(1, activation='sigmoid'))
print(model.summary())
#Loading pretrained word embeddings and Freezing the Embedding layer
model.layers[0].set_weights([embedding_matrix])
model.layers[0].trainable = False
# compile network
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
# fit network
history=model.fit(X_train, y_train, epochs=6,verbose=2)
# evaluate
loss, acc = model.evaluate(X_test, y_test, verbose=0)
print('Test Accuracy: %f' % (acc*100))
However, I am getting this error:
Traceback (most recent call last):
File "D:/banana.py", line 177, in <module>
history=model.fit(X_train, y_train, epochs=6,verbose=2)
File "D:\ProgramData\Miniconda3\envs\Env_DLexp1\lib\site-packages\keras\engine\training.py", line 1154, in fit
batch_size=batch_size)
File "D:\ProgramData\Miniconda3\envs\Env_DLexp1\lib\site-packages\keras\engine\training.py", line 642, in _standardize_user_data
y, self._feed_loss_fns, feed_output_shapes)
File "D:\ProgramData\Miniconda3\envs\Env_DLexp1\lib\site-packages\keras\engine\training_utils.py", line 284, in check_loss_and_target_compatibility
' while using as loss `categorical_crossentropy`. '
ValueError: You are passing a target array of shape (3, 1) while using as loss `categorical_crossentropy`. `categorical_crossentropy` expects targets to be binary matrices (1s and 0s) of shape (samples, classes). If your targets are integer classes, you can convert them to the expected format via:
```
from keras.utils import to_categorical
y_binary = to_categorical(y_int)
```
Alternatively, you can use the loss function `sparse_categorical_crossentropy` instead, which does expect integer targets.
I tried everything the error message says, but to no avail. After some research I came to the conclusion that the model is not trying to predict multiple classes, that's why the categorical_crossentropy loss is not being accepted. I then realized that, if I changed it for binary cross-entropy the error goes away, which is really a confirmation that this is not working as a multiclass classification model.
What can I do to adjust my code to make it work as intended? Am I S*it out of luck and have to start a whole different project?
Any type of guidance will be of immense help for me and my mental health.
You should make two changes. First the number of neurons in the output of your network should match the number of classes, and use the softmax activation:
model.add(Dense(4, activation='softmax'))
Then you should use the sparse_categorical_crossentropy loss as you are not one-hot encoding the labels:
model.compile(loss='sparse_categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
Then the model should be able to train without errors.
I got a dataset with 6 datapoints +4 datapoints as labels, they asked to predict those 4 timesteps using the 6 datasteps.
can you please advise me what model and how should I use it , I though about some kind of RNN since there is time for each point.
Thanks!
These sort of problems where the predictions depend on the previous inputs are generally uses RNN networks(rnn, gru and lstm) as they retain the previous state information.
for deeper understanding:
https://colah.github.io/posts/2015-08-Understanding-LSTMs/
Please go through the comments as well I have written in the code.
from __future__ import absolute_import, division, print_function, unicode_literals
import tensorflow as tf
from tensorflow.keras import Model
import numpy as np
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense
from tensorflow.keras.layers import RNN, LSTM
"""
creating a toy dataset
lets use this below ```input_sequence``` as the sequence to make data points.
as per the question, we will use 6 points to predict next 4 points
"""
input_sequence = [1,2,3,4,5,6,7,8,9,10,1,2,3,4,5,6,7,8,9,10,1,2,3,4,5,6,7,8,9,10]
X_train = []
y_train = []
#first 6 points will be our input data points and next 4 points will be data label.
# so on we will shift by 1 and make such data points and label pairs
for i in range(len(input_sequence)-9):
X_train.append(input_sequence[i:i+6])
y_train.append(input_sequence[i+6:i+10])
X_train = np.array(X_train, dtype=np.float32)
y_train = np.array(y_train, dtype=np.int32)))
#X_test for the predictions (contains 6 points)
X_test = np.array([[8,9,10,1,2,3]],dtype=np.float32)
print(X_train.shape)
print(y_train.shape)
print(X_test.shape)
#we will be using basic LSTM, which accepts input in ```[num_inputs, time_steps, data_points], therefore reshaping as per that```
X_train = np.reshape(X_train, (X_train.shape[0], 1, X_train.shape[1]))
X_test = np.reshape(X_test, (X_test.shape[0], 1, X_test.shape[1]))
print(X_train.shape)
print(y_train.shape)
print(X_test.shape)
x_points = X_train.shape[-1]
print("one input contains {} points".format(x_points))
model = Sequential()
model.add(LSTM(4, input_shape=(1, x_points)))
model.add(Dense(4))
model.compile(loss='mean_squared_error', optimizer='adam')
model.summary()
model.fit(X_train, y_train, epochs=500, batch_size=5, verbose=2)
output = list(map(np.ceil, model.predict(X_test)))
print(output)
we have used the simpler model, this further can be improved to get better results.
I am trying, just for practising with Keras, to train a network to learn a very easy function.
The input of the network is 2Dimensional . The output is one dimensional.
The function can indeed represented with an image, and the same is for the approximate function.
At the moment I'm not looking for any good generalization, I just want that the network is at least good in representing the training set.
Here I place my code:
import matplotlib.pyplot as plt
import numpy as np
from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation
from keras.optimizers import SGD
import random as rnd
import math
m = [
[1,1,1,1,0,0,0,0,1,1],
[1,1,0,0,0,0,0,0,1,1],
[1,0,0,0,1,1,0,1,0,0],
[1,0,0,1,0,0,0,0,0,0],
[0,0,0,0,1,1,0,0,0,0],
[0,0,0,0,1,1,0,0,0,0],
[0,0,0,0,0,0,1,0,0,1],
[0,0,1,0,1,1,0,0,0,1],
[1,1,0,0,0,0,0,0,1,1],
[1,1,0,0,0,0,1,1,1,1]] #A representation of the function that I would like to approximize
matrix = np.matrix(m)
evaluation = np.zeros((100,100))
x_train = np.zeros((10000,2))
y_train = np.zeros((10000,1))
for x in range(0,100):
for y in range(0,100):
x_train[x+100*y,0] = x/100. #I normilize the input of the function, between [0,1)
x_train[x+100*y,1] = y/100.
y_train[x+100*y,0] = matrix[int(x/10),int(y/10)] +0.0
#Here I show graphically what I would like to have
plt.matshow(matrix, interpolation='nearest', cmap=plt.cm.ocean, extent=(0,1,0,1))
#Here I built the model
model = Sequential()
model.add(Dense(20, input_dim=2, init='uniform'))
model.add(Activation('tanh'))
model.add(Dense(1, init='uniform'))
model.add(Activation('sigmoid'))
#Here I train it
sgd = SGD(lr=0.5)
model.compile(loss='mean_squared_error', optimizer=sgd)
model.fit(x_train, y_train,
nb_epoch=100,
batch_size=100,
show_accuracy=True)
#Here (I'm not sure), I'm using the network over the given example
x = model.predict(x_train,batch_size=1)
#Here I show the approximated function
print x
print x_train
for i in range(0, 10000):
evaluation[int(x_train[i,0]*100),int(x_train[i,1]*100)] = x[i]
plt.matshow(evaluation, interpolation='nearest', cmap=plt.cm.ocean, extent=(0,1,0,1))
plt.colorbar()
plt.show()
As you can see, the two function are completely different, and I can't understand why.
I think that maybe model.predict doesn't work as I axpect.
Your understanding is correct; it's just a question of hyperparameter tuning.
I just tried your code, and it looks like you're not giving your training enough time:
Look at the loss, under 100 epochs, it's stuck at around 0.23. But try using the 'adam' otimizer instead of SGD, and increase the number of epochs up to 10,000: the loss now decreases down to 0.09 and your picture looks much better.
If it's still not precise enough for you, you may also want to try increasing the number of parameters: just add a few layers; this will make overfitting much easier ! :-)
I have changed just your network structure and added a training dataset. The loss decreases down to 0.01.
# -*- coding: utf-8 -*-
"""
Created on Thu Mar 16 15:26:52 2017
#author: Administrator
"""
import matplotlib.pyplot as plt
import numpy as np
from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation
from keras.optimizers import SGD
import random as rnd
import math
from keras.optimizers import Adam,SGD
m = [
[1,1,1,1,0,0,0,0,1,1],
[1,1,0,0,0,0,0,0,1,1],
[1,0,0,0,1,1,0,1,0,0],
[1,0,0,1,0,0,0,0,0,0],
[0,0,0,0,1,1,0,0,0,0],
[0,0,0,0,1,1,0,0,0,0],
[0,0,0,0,0,0,1,0,0,1],
[0,0,1,0,1,1,0,0,0,1],
[1,1,0,0,0,0,0,0,1,1],
[1,1,0,0,0,0,1,1,1,1]] #A representation of the function that I would like to approximize
matrix = np.matrix(m)
evaluation = np.zeros((1000,1000))
x_train = np.zeros((1000000,2))
y_train = np.zeros((1000000,1))
for x in range(0,1000):
for y in range(0,1000):
x_train[x+1000*y,0] = x/1000. #I normilize the input of the function, between [0,1)
x_train[x+1000*y,1] = y/1000.
y_train[x+1000*y,0] = matrix[int(x/100),int(y/100)] +0.0
#Here I show graphically what I would like to have
plt.matshow(matrix, interpolation='nearest', cmap=plt.cm.ocean, extent=(0,1,0,1))
#Here I built the model
model = Sequential()
model.add(Dense(50, input_dim=2, init='uniform'))## init是关键字，’uniform’表示用均匀分布去初始化权重
model.add(Activation('tanh'))
model.add(Dense(20, init='uniform'))
model.add(Activation('tanh'))
model.add(Dense(1, init='uniform'))
model.add(Activation('sigmoid'))
#Here I train it
#sgd = SGD(lr=0.01)
adam = Adam(lr = 0.01)
model.compile(loss='mean_squared_error', optimizer=adam)
model.fit(x_train, y_train,
nb_epoch=100,
batch_size=100,
show_accuracy=True)
#Here (I'm not sure), I'm using the network over the given example
x = model.predict(x_train,batch_size=1)
#Here I show the approximated function
print (x)
print (x_train)
for i in range(0, 1000000):
evaluation[int(x_train[i,0]*1000),int(x_train[i,1]*1000)] = x[i]
plt.matshow(evaluation, interpolation='nearest', cmap=plt.cm.ocean, extent=(0,1,0,1))
plt.colorbar()
plt.show()